EP0368951B1

EP0368951B1 - Sparse superlattice signal processor

Info

Publication number: EP0368951B1
Application number: EP89904408A
Authority: EP
Inventors: George Carayannis; Christos Halkias; Dimitris Manolakis; Elias Koukoutsis
Original assignee: Adler Research Associates
Current assignee: Adler Research Associates
Priority date: 1988-03-11
Filing date: 1989-03-03
Publication date: 1994-02-16
Anticipated expiration: 2009-03-03
Also published as: DE68913151D1; DE68913151T2; JPH02503501A; CA1304822C; IL89505A0; IL89505A; EP0368951A1; WO1989008888A1

Abstract

An optimal parametric signal processing structure, which receives autocorrelation coefficients, produces, in an additive sparse superlattice embodiment, quantities lambdai+, defined as (1 - km)(km-1 + 1), and in a subtractive sparse superlattice embodiment, quantities lambdam-, defined as (1 + km)(1 - km-1), where the quantities ki correspond to the lattice predictor coefficients, or PARCORS of the signal under study. The additive and subtractive sparse superlattice processing structures are strictly optimal in terms of hardware and complexity, yet lend themselves to implementation in a fully parallel, fully sequential, or parallel partitioned manner.

Description

The present invention is related to the following patent applications:

WO-A-87/05 422; WO-A-87/05 421; EP-A-0 282 799 EP-A-0 303 051.

The entire disclosures of each of the four applications identified above are hereby incorporated by reference thereto.

FIELD OF THE INVENTION

The present invention relates to the field of signal processing, and in particular, the parametric signal processing.

BACKGROUND OF THE INVENTION

Parametric signal processing is used in many areas such as speech and image analysis, synthesis and recognition, neurophysics, geophysics, array processing, computerized tomography, communications and astronomy, to name but a few.
One example of signal processing of particular importance is the linear prediction technique which may be used for spectral estimation, and in particular for speech analysis, synthesis and recognition and for the processing of seismic signals, to enable the reconstruction of geophysical substrata. The linear prediction technique employs a specialized autocorrelation function.
Another form of signal processing which finds a multitude of applications, is the determination of an optimal (in the least square sense) finite impulse response (FIR) filter. A signal processor employing such a technique works with the autocorrelation of the filter input signal and the cross-correlation between the input and the desired response signal, and may be used in many of the abovementioned applications.
Still another form of signal processing of particular importance is known in the art as "L-step ahead" prediction and filtering, for solving the "optimum lag" problem. This technique is especially useful in designing spiking and shaping filters. Signal processors which perform this function employ a specialized autocorrelation function which also takes into account a time lag associated with the system.
Generally, as the order of the system under investigation increases, the complexity of the signal processing necessary to provide useful information also increases. For example, using the general Gaussian elimination procedure, a system of order p can be processed in 110(p3)" steps, indicating the number of steps as being "on the order of" p³, i.e., a function of p cubed. Thus, it will be appreciated that a system having order of p=100 requires on the order of one million processing steps to process the signal, a limitation of readily apparent significance, especially where real time processing is required.
Signal processing techniques have been developed which have reduced the number of operations required to process a signal. One such method has been based on a technique developed by N. Levinson, which requires 0(p²) sequential operations to process the signal. In particular, the "Levinson technique" requires 0(2·p²) sequential operations in order to process the signal. A version of this technique, known as the "Levinson-Durbin" technique, which is specialized for the linear prediction problem, requires 0(1 p2) sequential operations to process the signal. Neither of these schemes is suitable for parallel implementation. On the general subject of the Levinson and Levinson-Durbin techniques, see N. Levinson, "The Wiener RMS (Root-Mean-Square) Error Criterion in Filter Design and Prediction", J. Math Phys., Vol. 25, pages 261-278, January 1947; and J. Durbin, "The Filtering of Time Series Models", Rev. Int. Statist. Inst., Vol. 28, pages 233-244, 1960.
Although they represent an order of magnitude improvement over the Gaussian elimination technique, the Levinson and Levinson-Durbin techniques are too slow for many complex systems where real time processing is required.
Another way of implementing the main recursion of the Levinson-Durbin technique, for the computation of what is widely known as "lattice coefficients", was developed by Schur in 1917, in order to establish a system stability criterion. See I. Schur, "Uber Potenzreihen Die In Innern Des Einheitskreises Beschrankt Sind", J. Reine Angewandte Mathematik, Vol. 147, 1917, pages 205-232. Lev-Ari and Kailath, of Stanford University, have developed a different approach, based on the Schur and Levinson techniques, which provides a triangular "ladder" structure for signal processing. The Lev-Ari and Kailath technique uses the signal, per se, as the input to the processor, rather than autocorrelation coefficients, and it is used in the signal modelling context. See H. Lev-Ari and T. Kailath, "Schur and Levinson Algorithms for Non-Stationary Processes", IEEE International Conference on Acoustics, Speech and Signal Processing, 1981, pages 860-864.
In another modification to the Schur technique, Le Roux and C. Gueguen rederived the Schur algorithm, giving emphasis to the finite word length implementation, using fixed point arithmetics. See Le Roux and Gueguen, "A Fixed Point Computation of Partial Correlation, Coefficients", IEEE Transactions on Acoustics, Speech, and Signal Processing, June 1977, pages 257-259.
Finally, Kung and Hu, have developed a parallel scheme, based on the Schur technique, which uses a plurality of parallel processors, to process a signal, having order p, in O(p) operations, a significant improvement compared to the Levinson-Durbin technique. See Kung and Hu, "A Highly Concurrent Algorithm and Pipelined Architecture for Solving Toeplitz Systems", IEEE Transactions on Acoustics, Speech and Signal Processing, Vol. ASSP-31, No. 1, February 1983, pp. 66-76. However, the application of the Kung and Hu technique is severely limited insofar as it requires that the number of processors be equal to the order of the system to be solved. Thus, the Kung and Hu technique cannot process a signal produced by a system having an order greater than the number of parallel processors. System complexity is therefore a major limiting factor in using the Kung and Hu technique, insofar as many complex systems may have orders much higher than the number of parallel processors currently available in modern VLSI or other technology.
The co-pending patent applications, identified above, overcome the difficulties associated with the prior art signal processors and method by providing optimal techniques for obtaining lattice predictor coefficients (in WO-A-87/05 422 and WO-A-87/05 421), for obtaining lattice filter coefficients, for LS-FIR filtering (in EP-A-0 282 799), and for obtaining direct predictor and filter coefficients, for linear prediction and LS-FIR filtering (in EP-A-0 303 051
The techniques employed in the co-pending applications use "superlattice" and "superladder" structures to implement the recursions disclosed in the applications. These superlattice and superladder structures are "two-order" structures, meaning that the basic computational elements of such structures produce intermediate and final filter, or predictor, values, e.g., of order m + 1, by processing only values having order m. For example, FIG. 1 herein, which is based on FIG. 3 of WO-A-87/05 422 and WO-A-87/05 421, illustrate that the intermediate and final values f having orders m are produced from intermediate values of
having order m-1.
Specifically, a system signal is applied to a digital autocorrelator 10, FIG. 1, which produces autocorrelation coefficients ro through r₈, which characterize the system. The coefficients are applied to an input device 12, such as a digital register or memory. Each autocorrelation coefficient, except for the first, ro, and the last, r₈, is multiplied by a first lattice predictor coefficient k₁, which is produced from r_o and r_i ( and ⁰ ₁ ), according to the general formula:
The product of each such multiplication is added, individually, to the adjacent two autocorrelation coefficients, to produce the first intermediate values ¹ _n , where n=0, 2 through 8 and -1 to -6. For example, autocorrelation coefficient r₃, designated as ⁰ ₃ and ⁰ ₃ , for the sake of conformance with the intermediate variables, is multiplied by lattice predictor coefficient k₁ in multipliers 14 and 16, respectively, and autocorrelation coefficients r₂ and r₄ are added, individually, to each product, in adders 18 and 20, respectively, to produce a pair of first intermediate values ¹ _-2 and ₄ ¹, respectively. Similarly, another two first intermediate values, e.g., ¹ _-1 and ¹ ₃, are produced by multiplying autocorrelation coefficient r₂ by the lattice predictor coefficient ki , and adding, individually, the adjacent autocorrelation coefficients, ri and r₃ to the products.
The second intermediate values are derived from the first intermediate values in a similar manner. First, k₂ is derived from the ratio of ₂ and ¹ ₀ in accordance with the formula given above. Then, second intermediate values ² ₃ and ² ₀, for example, are produced by multiplying the first intermediate values ¹ _-1 and ?¹ ₂ by the lattice predictor coefficient k₂, and adding the adjacent first intermediate values ¹ ₃ and ¹ ₀ to the products, individually. The signal processing continues according to the basic lattice recursion disclosed in WO-A-87/05 422 and WO-A-87/05 421 until the final values ⁷ ₈ and ⁷ ₀ are obtained, from which the last lattice predictor coefficient k₈ can be produced.
The lattice predictor coefficients k_; completely characterize the linear predictor and can be used instead of the direct predictor coefficients a. They are many times preferred for storage, transmission and fast speech synthesis, since they have the advantages of being ordered, bounded by unity and can readily be used for stability control, efficient quantization, and the like. Since r_o corresponds to the energy of the signal, and will therefore have the greatest amplitude of any of the signals processed by the superlattice, all variables can be normalized with respect to ro, thus facilitating the use of fixed point processing, with its advantages of precision, speed and processing simplicity.
FIG. 2 illustrates a butterfly basic cell from which the superlattice of FIG. 1 can be constructed, as disclosed in WO-A-87/05 422. As taught by that co-pending application, the superlattice structure of FIG. 1 can be implemented by the repeated use of such butterfly basic cells, in the manner illustrated in FIG. 3. The reference numerals used in FIG. 2 correspond to those of FIG. 1.
Referring to FIG. 2, it will be appreciated that the multiplication of r₃ by k₁ is performed twice, in multipliers 14 and 16. This double operation can be eliminated by employing a single multiplier 16' in a "reduced basic cell", illustrated in FIG. 4. In this case, the multiplier 16' provides two outputs, one of which is added to r₂ in adder 18, to provide the intermediate quantity 2 , and a second output which is added to r₄ in adder 20, to produce intermediate value ¹.This hardware reduction, however, can only be achieved for those basic cells used to produce the first intermediate values, since all succeeding basic cells receive at their inputs four distinct quantities.

OBJECTS AND SUMMARY OF THE INVENTION

It is therefore an object of the present invention to provide an optimal parametric signal processing structure in which such double operations are eliminated throughout the processing structure.
It is a further object of the present invention to provide an optimal parametric signal processor of strictly minimal hardware complexity.
It is a further object of the present invention to provide an optimal signal processor, having minimal hardware complexity, which can be implemented using (i) a single processing unit to process a signal in a fully sequential manner, (ii) a plurality of processing units to process a signal in a fully parallel manner, or (iii) a lesser plurality of processing units to process a signal in a partitioned parallel manner.
It is a further object of the present invention to provide an optimal parametric signal processor which employs a three-order structure as a basic cell.
It is a further object of the present invention to provide an optimal parametric signal processor which is minimal in terms of hardware complexity, and which permits a finite word length implementation.
It is a further object of the present invention to provide an optimal parametric signal processing structure, having minimal hardware complexity, for producing (i) lattice predictor coefficients, (ii) linear predictor coefficients, (iii) linear phase predictor coefficients, (iv) constant group delay predictor coefficients, and (v) LS-FIR optimal filter coefficients.
It is yet another object of the present invention to provide a structure called the "additive sparse superlattice", which receives autocorrelation coefficients, and produces therefrom quantities λ_i ⁺, defined as (1 - k_i) (k_i-1 + 1), where the quantities k_; correspond to the lattice predictor coefficients, or "PARCORS" of the signal.
A particular implementation of the additive sparse superlattice includes a plurality of adders, a plurality of multipliers, a storage and retrieval structure, and a divider. The storage and retrieval structure selectively applies a first plurality of pairs of adjacent autocorrelation coefficients to a plurality of adders to produce an associated plurality of first intermediate values. The storage and retrieval structure selectively applies a plurality of the autocorrelation coefficients to the plurality of multipliers to produce an associated plurality of first products, and also selectively reapplies an adjacent pair of the first intermediate values, and a selected one of the first products, to each of a plurality of the adders, to thereby produce an associated plurality of second intermediate values. The divider initially divides a selected first intermediate value by a value related to the zeroth order autocorrelation coefficient, to thereby produce a first of the quantities λ_i+, and thereafter divides a selected second intermediate value by the selected first intermediate value, to thereby produce a second of the quantities λ_i+. The first of the quantities λ_i+ is applied to the plurality of multipliers to produce the plurality of second intermediate values.
It is yet another object of the present invention to provide a structure called the "subtractive sparse superlattice" which receives autocorrelation coefficients and produces therefrom quantities λ_i ^-, defined as (1 + k_m) (1 - k_m-₁), where the quantities k; correspond to the lattice predictor coefficients, or PARCORS of the signal.
A particulr implementation of the subtractive sparse superlattice also includes a plurality of adders, a plurality of multipliers, a storage and retrieval structure, and a divider. The storage and retrieval structure selectively applies a plurality of pairs of differences between adjacent autocorrelation coefficients to the plurality of adders, to produce an associated plurality of second intermediate values. The storage and retrieval structure also selectively applies a plurality of differences between adjacent autocorrelation coefficients to the plurality of multipliers to produce an associated plurality of first products, and selectively reapplies an adjacent pair of the second intermediate values, and a selected one of the first products, to each of a plurality of the adders, to produce an associated plurality of third intermediate values. The divider initially divides a selected second intermediate value by the difference between the zeroth and first autocorrelation coefficients, to thereby produce a first of the quantities X_i-, and thereafter divides a selected third intermediate value by the selected second intermediate value, to thereby produce a second of said quantities X_i ^l. The first of the quantities X_i ^l is applied to the plurality of multipliers to produce the plurality of third intermediate values.

BRIEF DESCRIPTION OF THE DRAWINGS

These and other objects, aspects and embodiments of the present invention will now be described in more detail with reference to the following drawing figures of which:

FIG. 1 is an illustration of a superlattice processing structure, in accordance with WO-A-87/05 422 and WO-A-87/05 421;
FIG. 2 is an illustration of the "butterfly basic cell" which can be employed to produce the superlattice processing structure of FIG. 1;
FIG. 3 is a diagram illustrating how the superlattice structure can be implemented through the repeated use of the butterfly basic cell of FIG. 2;
FIG. 4 is an illustration of a "reduced basic cell" which reduces the redundancies inherent in the butterfly basic cell structure of FIG. 2;
FIGS. 5A and 5B respectively illustrate "sparse basic cells" for the additive and subtractive sparse superlattice structures, in accordance with the present invention;
FIG. 6 illustrates the additive sparse superlattice processing structure in accordance with a first aspect of the present invention;
FIG. 7 illustrates the subtractive sparse superlattice processing structure in accordance with another aspect of the present invention;
FIG. 8 illustrates a parallel partitioned implementation of the subtractive sparse superlattice;
FIG. 9 is a diagram illustrating the signal flow through the subtractive sparse superlattice, using a partitioned partitioned implementation;
FIG. 10 is a schematic diagram illustrating a hardware implementation of the subtractive sparse superlattice of FIGS. 8 and 9;
FIG. 11 is an illustration of a side-fed sparse superlattice, for use with the additive sparse superlattice, for producing a linear phase predictor;
FIG. 12 illustrates a processing structure for producing a linear predictor, based on the production of the linear phase prediction coefficients by the structure of FIG. 11;
FIG. 13 is an illustration of a side-fed sparse superlattice, for use with the subtractive sparse superlattice, for producing the constant group delay predictor; and
FIG. 14 is an illustration of a sparse superladder processing structure for providing optimal LS-FIR filtering.

DETAILED DESCRIPTION OF THE INVENTION

The substitution of a "reduced basic cell", involving one multiplication, for the "butterfly" or "lattice" basic cells, which involve two multiplications, would be desireable, but it is apparent from the foregoing discussion of FIG. 4 that even with an intuitive belief that such a reduction to the superlattice of FIG. 1 may be possible, the implementation of such a reduction, by a simple topological observation of the superlattice, is not straightforward. The following analysis provides an analytical tool to illustrate how the communication between "reduced basic cells", within a superlattice, is organized, and how the recovery of the PARCORS from reduced basic cells is possible. The analysis will begin with an example. Let us define the value
as the sum of two adjacent intermediate values of order j, in the superlattice of FIG. 1. For example:
Using the lattice recursions for he ^j _i , we can write:

Consequently:
Defining:

equation (4) can be written:
Since:

equation (7) can be written:
where:
This result can be generalized independently of the order selected.
By defining:
the following relationship holds:
Using the basic lattice recursion (or the Schur recursive formula) each one of the quantities on the right hand side of (12) can be written:

or:
Consequently:
or:
But:
and:
Substituting into equation (17) we obtain:
But:
Therefore:
Defining:
equation (22) can be written:
FIG. 5A is an illustration of an additive "sparse basic cell" ("SBC") which implements the above relationship. Specifically, the quantity
is multiplied by -λ⁺ _m in multiplier 30, and the product is added to the quantities ψ_i ^m and
in adder 32, to produce the output
The SBC of FIG. 5A is called "additive", and the quantities are given a "+" superscript, to indicate that the output of the SBC represents the sum of the intermediate values of the superlattice of FIG. 1.
A complete processing structure, which will be called the "sparse superlattice" can be constructed with the repeated use of the SBC shown in FIG. 5A. Before discussing the sparse superlattice, however, the following will demonstrate that a structure operating with the sums of the intermediate values of the superlattice of FIG. 1 can provide information relating to the lattice predictor coefficients.
We will define X^ψ ₀₄ as follows:
since:
Equation (25) represents the well known total squared error computation in linear prediction.
From equation (24), generalized for an arbitrary order m, k_m can be produced as follows:
Therefore, a recursive relationship exists for the processing of the lattice predictor coefficients, even when only the sums of the diate
Equation (24), generalized for an arbitrary order m, also gives the following definition of λ⁺ _m :
Thus, the quantities λ+m are readily available, in a straightforward manner, from the quantities ψ^m ₀ and
The sparse basic cell of FIG. 5A can be used as the basic building block for what will be called the "additive sparse superlattice", illustrated in FIG. 6. This structure can be used for linear prediction, and with additional circuitry, the lattice predictor coefficients can be produced. Also, the linear prediction coefficients can be obtained either from the lattice prediction coefficients, or from the quantities λ_i ⁺, with additional circuitry, described below.
With specific reference to FIG. 6, the autocorrelation coefficients ro to r₇, for a system having order 7, are applied to the inputs 50-57 of the sparse superlattice, as shown. Since the sums of quantities are processed in the sparse superlattice, the sums of the autocorrelation coefficients are shown as being applied to the sparse superlattice, appropriately divided by two. For example, autocorrelation coefficient r_o is applied to input 50, but is shown as the quantity (r₀ +r₀)/². Similarly, autocorrelation coefficients r_i through r₇ are applied to inputs 51-57, respectively. With the exception of autocorrelation coefficients ro and r₇, each of the autocorrelation coefficients is applied to two adders (designated in the figures by circles) and one multiplier (designated in the figures by triangles); autocorrelation coefficient ri is applied to adders 60 and 61; autocorrelation coefficient r₂ is applied to adders 61 and 62; autocorrelation coefficient r₃ is applied to adders 62 and 63, autocorrelation coefficient r₄ to adders 63 and 64, and so on. Autocorrelation coefficients ri through r₆ are also applied to multipliers 71 through 76, respectively. Finally, autocorrelation coefficient ro is applied to adder 60, and autocorrelation coefficient r₇ is applied to adder 66. The sums produced at adders 60 through 66 yield first intermediate values ψ¹ _i, where i = 0 through-6.
Dividers 81 through 87 function to produce the quantities -λ⁺m in accordance with equation (27), above. Autocorrelation coefficient r₀/², shown as ψ₀ ⁰, for the sake of conformance with the intermediate values, is applied to the divider 81 along with the intermediate value ψ¹ ₀ , to produce the quantity -X_i ^+. Similarly, the intermediate value ψ² ₀ is divided by the intermediate value ψ¹ ₀ in divider 82, to produce the quantity -λ⁺ ₂ . The quantities -λ⁺ ₃ through -λ⁺ ₇; are produced in a similar manner, by dividers 83 through 87, respectively.
The quantities -λ⁺ _m produced by the dividers 81-87, are applied to associated multipliers in accordance with the basic recursion of equation (23). Specifically, the quantity -X+ , is applied to multipliers 71 through 76, the quantity -λ⁺ ₂ is applied to multipliers 101 through 105, and so on, as shown in FIG. 6. Each of the multipliers shown in FIG. 6 functions to multiply the input value by a respective one of the quantities -λ⁺ _m . For example, the value ψ⁰-1 , is multiplied by -X_i ⁺, in multiplier 71, ψ¹-1 , is multiplied by -À⁺ ₂ in multiplier 101, and so on, as shown.
The output of each of the multipliers is added to the contents of the two adjacent adders, to produce the next order of intermediate values _ψij. For example, the output of multiplier 71 is added to the outputs of adjacent adders 60 and 61, in adder 91, to produce a second intermediate value _Ψ ⁰ ₂ ; the output of multiplier 101 is added to the output of adjacent adders 91 and 92, in adder 99, to produce a third intermediate value ψ ³0
Adders 92 through 96 produce second intermediate variables ψ²i, where i=0 through -5, and the remaining intermediate values are produced as shown. In this manner, the additive sparse superlattice of FIG. 6 implements the basic additive sparse recursion of equation (23), using the additive sparse basic cell of FIG. 5A as a basic building block.
It will be noted that each of the intermediate values is related to the intermediate values of the original superlattice, shown in FIG. 1, by the following equation:
It should also be noted that unlike the original superlattice of FIG. 1, each of the sparse basic cells of FIG. 5A is a "three order" structure, insofar as intermediate values having orders m-1, m and m + 1 are involved in each basic cell. For example, the basic cell which includes multiplier 103 receives a first intermediate value from adder 63, and two second intermediate variables, in adders 93 and 94, to produce a third intermediate variable ψ³-₂.
Since the double operations inherent in the superlattice of FIG. 1 do not exist, the sparse superlattice is minimal in terms of hardware complexity. For example, comparing the prior superlattice structure of FIG. 1 to the additive sparse superlattice structure of FIG. 6, it will be seen that the implementation of the superlattice structure of FIG. 1, for a system having order 7 (reduced from order 8 as shown in FIG. 1) requires some 42 multiplications and 42 additions, while the implementation of the additive sparse superlattice shown in FIG. 6 requires only 28 additions and 21 multiplications.
As noted above, the intermediate values associated with the additive sparse superlattice of FIG. 6 consist of the sums of the intermediate values of the original superlattice, as set forth in equation (28), and the sparse superlattice of FIG. 6 is therefore called "additive". However, a subtractive sparse superlattice, in which the intermediate values consist of the differences between the intermediate values of the original superlattice, also exists. A subtractive sparse superlattice, for processing a signal having order 10, is shown in FIG. 7, and employs the subtractive SBC of FIG. 5B which is structurally the same as the additive SBC of FIG. 5A, but which receives different quantities.
In the subtractive sparse superlattice, the intermediate values
are defined as follows:
The basic subtractive sparse recursion is as follows:
where:
The structure shown in FIG. 7 is simplified over that of FIG. 6 due to the absence of the first coefficient λ₁. This can be seen with the example of φ-2²:
Thus, unlike the additive sparse superlattice, there is no need to multiply the autocorrelation coefficients by the value λ₁ in order to obtain the first intermediate values φ_i ¹.
In FIG. 7, the differences between adjacent autocorrelation coefficients are applied to inputs 110-119. Specifically the value of r_o minus r_i, corresponding to the first intermediate value φ¹ ₀ , is applied to input 110, the value of ri minus r₂, corresponding to the first intermediate value φ1-1, is applied to input 111, the value of r₂ minus r₃, corresponding to first intermediate value φ1-2, is applied to input 112, and so on, such that the remaining first intermediate values φ_i ¹, where i equals -3 to -9, are applied to inputs 113 through 119, respectively. The first intermediate values φ¹_1 , through φ¹-8 are each applied to a multiplier and a pair of adders on either side of the multiplier. For example, the first intermediate value at input 111 is applied to multiplier 131 and to two adders 120 and 121, on either side of multiplier 131. Similarly, the value applied to input 112 is applied to multiplier 132 and to adders 121 and 122, on either side of multiplier 132, and so on. The values at the bounding inputs 110 and 119 are applied to single adders 120 and 128, respectively. At this point, the second intermediate values φ_i ², for i=0 through -9, are produced in adders 120-128, respectively.
The quantities
are produced in accordance with Equation (31) above. For example, the quantity -X2 is produced by dividing, in divider 141, the second intermediate value φ ²0 , from adder 120, by intermediate value φ¹ ₀, from input 110. In a manner similar to that described with reference to the additive sparse superlattice of FIG. 6, the remaining dividers 142-149 function to divide the intermediate values to produce an associated quantity
in accordance with the above relationship.
Also in a manner similar to that described with reference to FIG. 6, the quantities -\m , produced by the dividers 141-149, are applied to associated multipliers, in accordance with the basic recursion of equation (30). Specifically the quantity -À2 is applied to multipliers 131-138, the quantity -λ-3 is applied to multipliers 161-167, and so on, as shown in FIG. 7. Each of the multipliers shown in FIG. 7 functions to multiply the input value by a respective one of the quantities -λm .
For example, the value φ1-1 is multiplied by -λ 2 in multiplier 131, φ¹-2 is multiplied by -X2 in multiplier 132, and so on, as shown.
The output of each of the multipliers in FIG. 7, is added to the contents of the two adjacent adders, to produce the next order of intermediate values φ_ij . For example, the second intermediate values in adders 121 and 122 are added to the output of multiplier 132, in adder 152, to produce the third intermediate value φ3 -1, in accordance with the basic subtractive sparse recursion of equation (30) above. Similarly, the second intermediate values on either side of their associated multipliers 131-138 are added to the output of the associated multiplier, in adders 151-158, to produce respective third intermediate values φ_j3, where i = 0, through -7. This process is continued, as shown, to produce the remaining intermediate values and quantities λi-.
The Table below summarizes the basic relationships governing both the original superlattice and the additive and subtractive sparse superlattices:
The additive and subtractive sparse superlattices introduce the new parameters λ_i+ and λi; , which are related to the lattice predictor coefficients k_i and the linear predictor coefficients a_i. Therefore, if λ_i+ or λi are produced, the lattice and/or linear predictor coefficients k_i and a_i can also be produced. Like k_i, which has an absolute value bounded by unity, the values of λ_i +and λi are bounded by the values 0 and 4. That is:
The sparse superlattices of FIGS. 6 and 7 illustrate the processing structures involved in producing the quantities λ_i+ and λ_i . As implemented, however, the entire structures illustrated therein do not necessarily exist, as a whole, at any particular point in time. Rather, because of the recursive nature of the sparse superlattices, the various adders, multipliers and dividers can be implemented by a relatively small number of such elements, which are shared in time, the outputs of the elements being recursively fed back to the appropriate inputs thereof, to produce the entire superlattice structures.
Like the original superlattice structures disclosed in the four copending applications, identified above, the additive and the subtractive sparse superlattices can be implemented, in this recursive manner, in a fully parallel, fully sequential or partitioned parallel implementation. In the fully parallel implementation, in the system of FIG. 7, a total of nine adders and eight multipliers would be provided, corresponding to adders 120 through 128 and multipliers 131 through 138, so that the quantities applied to the inputs 110 through 119 could be processed, in parallel, to produce the second and third intermediate values, which would be fed back to a subset of the same adders and multipliers, to recursively produce the fourth intermediate values, and so on.
In the fully sequential implementation, a single multiplier, and at least one adder would be provided, and the second intermediate values would be produced sequentially, until all the second intermediate values were produced. The multiplier then would function to produce the third intermediate values, one after another, and so on, until all intermediate values were produced.
FIG. 8 is an illustration of the processing performed by the partitioned parallel implementation of the subtractive sparse superlattice. The partitioned parallel implementation was chosen for this example insofar as it combines the elements of the fully parallel and the fully sequential implementations. In view of the following discussion, and specific example, the fully parallel and fully sequential implementations will be readily apparent to those skilled in the art.
With reference to FIG. 8, three multipliers 131, 132 and 133, and three adders 121, 122 and 123 are provided. Initialization is first performed to determine the second intermediate value φ² ₀ and the value of -X-₂ by adding the values of φ ¹0 and φ¹_1 in adder 120 (which, in actuality, is one of the adders 121 through 123, as explained below), and by dividing the sum by φ ¹0 in divider 141. The intermediate value φ² ₀ is stored for later use in the first partition. The first intermediate values φ¹_1, φ¹-2 and φ1-3 are then applied to the adders 121 through 123, to produce the second intermediate values φ²-1, φ²-2 and φ²-3. These first intermediate values are also applied to respective ones of the three multipliers 131 through 133, the outputs of which are then added to the second intermediate variables on either side of each respective multiplier, in adders 151-153 (which, in actuality, are adders 121-123, as explained below), to produce third intermediate values φ³ ₀ , φ³ _-1, and φ³-2. At this point, -À3 can be produced by divider 142, (which, like dividers 143 and 144, is in actuality divider 141). Two of the second intermediate values, λ² _-1 , and À^2- ₂, are applied to multipliers, and the respective products are added to the adjacent two third intermediate values to produce fourth intermediate variables φ⁴ ₀ and φ⁴ _-1, as shown. One of the third intermediate values, φ³ _-1 is then applied to one of the multipliers and the product is added to the two fourth intermediate variables to produce one of the fifth intermediate values φ⁵ ₀ . The intermediate values along the border of the first partition, namely φ² _-3 , φ³-2, φ⁴-1 and φ⁵-0, are stored for later use in the second partition.
At this point the first partition is completed, and the processing of the second partition begins by applying the first intermediate values φ¹_4 , φ¹-5 and φ¹-6 to the three adders, designated in FIG. 8 as 124, 125 and 126 (but which in actuality are adders 121-123) and to the three multipliers, designated in the figure as 134, 135 and 136, (but which in actuality correspond to multipliers 131-133). In this manner, the intermediate variables are generated through the second partition, the three adders and multipliers functioning to cut through the sparse superlattice in sequential partitions, to provide the processing structure of the entire sparse superlattice. The three adders and multipliers in this example function in parallel, but process each of the partitions sequentially. In this manner, any available number of adders and multipliers can be employed to process a signal having any order, the adders and multipliers functioning in parallel to process each partition, with the required number of partitions being performed in sequence, until the entire sparse superlattice is completed.
The signal flow through the sparse superlattice, using a parallel partition implementation, will be explained in more detail with reference to FIG. 9 which illustrates a sparse superlattice implementation for processing a signal having order 8. In this figure, three adders 121, 122 and 123, three multipliers 131, 132 and 133, and a divider 141, corresponding to the adders, multipliers and dividers of FIGS. 7 and 8 are shown, with an additional notation in parentheses following each reference number, indicating the number of times each element has been used, up to that point, within the processing structure. For example, reference numeral 122(2) refers to the second use of adder 122, reference numberal 122(3) refers to the third use of adder 122, and so on. This notation differs from hat of FIGS. 7 and 8 since the adders, multipliers and dividers shown in positions downstream of the adders 121 through 123, multipliers 131 through 133 and divider 141, do not actually exist, as noted above, but rather are implemented by the original elements, through the appropriate use of feedback, to be described in detail below.
Each of the adders 121-123 has three inputs, a, b and c, and an output d equal to the sum of the inputs. Each of the multipliers 131 - 133 has two inputs, λ and e, and an output f equal to the product of λ and e, and the divider 141 has two inputs g and h, and an output i equal to -h/g.
For initialization, the quantity r_o-r_i, herein defined as δr₁ , and the quantity r_i-r₂, defined as δr2, are applied to the a and c inputs of adder 121 (1), the middle input, b, being set to zero, to produce the quantity d, which is stored, in a bordering buffer, denoted in the figure by "bb1 ". The quantity δr1 is also applied to the g input of divider 141 (1), and the quantity d from adder 121 (1) is applied to the h input of the divider, to produce the quantity i, corresponding to -λ₂ . The quantity -λ₂ is stored in a λ buffer, for later use.
After initialization, adder 121 (2), in its second operation (indicated by the parenthetical 2 appended to the reference numeral) receives the quantity δr₂ at its a input and the quantity δr₃, corresponding to the quantity r₂-r₃, at its c input. As before, the b input is set to zero. Similarly, the quantity δr₃ is applied to the a input of adder 122(1) (which is performing its first operation) while the quantity δr₄, corresponding to r₃-r₄, is applied to the c input of adder 122(1), the b input being set to zero. Finally, the quantity sr4 is applied to the a input of adder 123(1) (performing its first operation), while its c input receives the quantity δr5, corresponding to the quantity r₄-r₅, the b input again being set to zero. The adders 121 (2), 122(1) and 123-(1) produce outputs d, which are appropriately fed back to the inputs of the adders and multipliers, the output of adder 123(1) also being stored in the bordering buffer, for later use in the second partition.
Simultaneously with the operations being performed by adders 121 (2), 122(1) and 123(1), the value of λ 2 is applied to the multipliers 131(1), 132(1) and 133(1), at the λ inputs thereto. Multiplier 131(1) receives the quantity δr₂, multiplier 132(1) receives δr₃ and multiplier 133(1) receives δr4, at their respective e inputs.
The d output of adder 121 (1), which was stored in a bordering buffer (bb₁is fed back to the a input of the same adder, denoted by 121 (3), the f output of multiplier 131 (1) is fed back to the b input of adder 121-(3), and the d output of the same adder 121 (2) is fed back to its c input, of adder 121 (3). The d output of adder 121 (2) is also applied to the a input of adder 122(2), the output of multiplier 132(1) is applied to the b input of adder 122(2), and the d output of 122(1) is fed back to its c input, of adder 122(2). Continuing, the d output of adder 122(1) is applied to the a input of adder 123(2), the f output of multiplier 133(1) is applied to the b input of adder 123(2), and the d output of adder 123(1) is fed back to its c input, of adder 123(2), which in turn produces an output d which is stored in the bordering buffer for use in the second partition.
At this point, the d output of adder 121 (3) is applied to the h input of divider 141 (2), and the d output from adder 121(1) is retrieved from storage and applied to the g input of divider 141 (2), to produce the quantity -λ 3 , which is stored in the λ buffer for later use. The quantity -λ 3 is then applied to only two multipliers 132(2) and 133(2), as the partition cuts through the sparse superlattice, as shown in FIG. 8. Likewise, only two adders 122(3) and 123(3) are employed to produce d outputs, the d output of adder 123-(3) being stored in the bordering buffer for later use, and the value -λ 4 , produced in divider 141 (3), is stored in the buffer. In the next iteration, only a single multiplier 133(4), and a single adder 123(4) is employed to produce a single d output, which is applied to the bordering buffer for later use, and to the divider 141 (4) for the production of the quantity λ 5 , which is stored in the buffer for later use. At this point, the first partition is complete.
The second partition begins with the application of δr₅ to the a input of adder 121 (4), and the application of δr6, corresponding to rs - r₆, to the c input of adder 121(4), the b input having been set to zero. Similarly, the quantities δr6 and δr7 are applied to the a and c inputs of adder 122(4), the quantities δr7 and δr8 are applied to the a and c inputs of adder 123(5), and the b inputs of adders 122(4) and 123(5) are set to zero. The adders 131(2), 132(3) and 133(5) receive the quantity λ 2 from the λ buffer, and the processing continues, as shown, until the second partition, and in the example shown in FIG. 9, the entire sparse superlattice, is complete.
One example of an implementation of the processing structure shown in FIG. 9, will now be described with reference to FIG. 10. Three adders 201-203 and three multipliers 204-206 are provided, and may be formed by three processors, each having an adder and multiplier, although such an arrangement is not required. Furthermore, even though the adders are shown as being separate, in a hardware sense, from the multipliers, a single hardware element may perform both addition and multiplication functions, if desired. Also, although implemented using three adders and multipliers, greater or lesser numbers of such elements can be used, as desired.
The adders 201-203, hereinafter called the "top", "middle" and "bottom" adders, respectively, each have three input registers, a, b and c, and function to add the contents of the input registers. The adders also have an output register d for storing the sum of the contents of the input registers. The multipliers 204-206, hereinafter called "top", "middle" and "bottom" multipliers, respectively, each have two input registers λ and e and function to multiply the contents of the input registers. The multipliers also have an output register f for storing the product.
A λ buffer 212 is provided for receiving, storing and transmitting values of λ_i, via λ buffer bus 210. Similarly, a or buffer 216 is provided for receiving, storing and transmitting values of δri, via or bus 214, and a bordering buffer 220 functions to receive, store and transmit intermediate values, via bordering buffer bus 218. Finally, a divider 222 has two input registers g and h, and output register, i, and functions to provide the quotient -h/g, and store it in the i register.
The λ buffer 212 receives as its only input the output of the i register in divider 222 by way of λ- bus 210. The λ buffer 212 provides an input to the registers of the multipliers 204-206. The or buffer receives as its inputs the quantities ro-r1, ri-r2, etc., as shown in FIG. 9, and provides inputs to the a, b and c registers of the adders 201-203, by way of the or bus 214. The bordering buffer 220 receives an input from the d register of the top adder 201, by way of bordering buffer bus 218, and provides inputs to the e register of the bottom multiplier 206, and to the a register of the bottom adder 203, also by way of the bordering buffer bus 218.
The output of the d registers of each of the adders 201 through 203 is applied as an input to its respective c register. The d register of the bottom adder 203 also provides an input to the a register of the middle adder 202, the e register of the middle multiplier 205, and the g and h registers of the divider 222. The d register of the middle adder 202 in a similar manner, provides an input to the a register of the top adder 201, the e register of the top multiplier 204, and to the g and h registers of the divider 222. The d register of the top adder 201 provides a third input to the g and h registers of the divider 222.
Finally, the f register of the bottom multiplier 206 provides an input to the b register of the bottom adder 203, the f register of the middle multiplier 205 provides an input to the b register of the middle adder 202, and the f register of the top multiplier 204 provides an input to the b register of the top adder 201. In describing the operation of the implementation shown in FIG. 13, the following notation will be used:
The operation of the implementation of FIG. 10 is as follows:

Initialization

Phase 1 : δr₂ - c(BA) 0 - b(BA) δr1 → (a(BA), g)
Phase 2: C [BA; 0, 1]
Phase 3 : d(BA) → (bb1, h)
Phase 4 : D [0,1]
Phase 5 : i → (λ₂ , g) Partition 1
Phase 6 : 0 → (b(BA), b(MA), b(TA)) δr₂ → (e(BM), a(BA)) δr₃ → (c(BA), e(MM), a(MA)) δr₄ → (c(MA), e (TM), a(TA)) δr₅ → c(TA)
Phase 7 : C [BA, MA, TA; 1, 1 ]
Phase 8 : d(BA) → (c(BA), a(MA))
d(MA) → (c(MA), a(TA)) d(TA) → (c(TA), bbi)
Phase 9 : λ₂ → (λ(BM)), X(MM), λ(TM))
Phase 10 : C [BM, MM, TM; 1,1]
Phase 11 : f(BM) → b(BA) f(MM) → b(MA) f(TM) → b(TA) d(BA) → e(MM) d(MA) → e(TM)
Phase 12 : C [BA, MA, TA; 1,2]
Phase 13 : d (BA) → (h, a(MA)) d(MA) → (c(MA), a(TA)) d(TA) → (c(TA), bb₂)
Phase 14 : D [1,2]
Phase 15 : i → (λ₃ , λ(MM), λ(TM))
Phase 16 : C [BM, MM; 1,2]
Phase 17 : f(MM) → b(MA) f(TM) → b(TA) d(MA) → e(TM)
Phase 18 : C [MA, TA; 1,3]
Phase 19 : d(MA) → (h, a(TA)) d(TA) → (C(TA), bb₃)
Phase 20 : D [1,3]
Phase 21 : i - (λ₄ , λ(TM))
Phase 22 : C [TM; 1,4]
Phase 24 : f(TM) - b(TA)
Phase 25 : C [TA; 1,4]
Phase 26 : d(TA) → (g, bb₄)
Phase 27 : D [1,4]
Phase 28 : i → (λ₅ , g) Partition 2
Phase 29 : δr₅ → (a(BA), e(BM)) δr₆ → (c(BA), a(MA), e(MM)) δr₇ → (c(MA), a(TA), e(TM)) δr₈ → c(TA)
0 → b(BA), b(MA), b(TA) Phase 30 : C [BA, MA, TA; 2,1 ]
Phase 31 : d(BA) → (c(BA), a(MA)) d(MA) → (c(MA), a(TA)) d(TA) → (c(TA), bb_T) bb₁ → a(BA) λ₂ → (λ(BM), λ(MM), λ(TM))
Phase 32 : C [BM, MM, TM; 2,1]
Phase 33 : f(BM) → b(BA) f(MM) → b(MA) f(TM) → b(TA) bb₁ → e(BM) d(BA) → e(MM) d(MA) → e(TM)
Phase 34 : C [BA, MA, TA; 2,2]
Phase 35 : d(BA) → (c(BA), a(MA)) d(MA) → (c(MA), a(TA)) d(TA) → (c(TA), bb₁ ) bb₂ → a(BA)

λ₃ → (λ(BM), λ(MM), λ(TM)) Those skilled in the art will appreciate the continuation of partition 2 in view of the above discussion. Also, although the above example implements the subtractive sparse superlattice, implementations for the additive sparse structure of FIG. 6 will be apparent to those skilled in the art, in view of the above discussion.
Once either the quantities λ_i or λ_i+ are obtained, the production of the lattice predictor coefficients (PARCOR's) is straightforward using the relevant relationships shown in the above table, which are reproduced here for the sake of convenience:

thus:
or:
Therefore, production of the lattice predictor coefficients can be achieved by using additional phases of the hardware illustrated in FIG. 10, or with additional hardware, if desired.
Other processes which can be implemented using one of the sparse superlattices are (i) the production of the linear prediction coefficients from the PARCOR's, (ii) the production of lattice filter coefficients, by employing the structures disclosed in EP-A-0 282 799, (iii) or the production of linear filter coefficients, by employing the structures disclosed in EP-A-0 303 051.
In addition to the above, the additive and subtractive sparse superlattice structures readily lend themselves to the production of linear phase predictors and constant group delay predictors. Specifically, a "unit side-fed" sparse superlattice, shown in FIG. 11, produces the linear phase predictor γ⁺ based on the quantities λ_i+produced from the additive sparse superlattice of FIG. 6, as follows:
The linear phase predictor has the following property:
so that for a predictor of order 7, as in FIG. 11,
The unit side-fed sparse superlattice of FIG. 11 applies unity to a series of side-fed inputs 301 through 308. Each of the unity inputs, except for the uppermost, 308, is applied to an associated multiplier 311 through 317, which functions to multiply the input by quantities -X+ , through
, respectively, which quantities were derived from the additive sparse superlattice.
The unity input 302 is multiplied by 2 in multiplier 318 and the result is added to the output of multiplier 311, in adder 320, to produce the first intermediate value γ+ 1,1. The unity input 303 is added to the output of multiplier 312, and to the first intermediate value γ+ 1,1, in adder 322, to produce a second intermediate value γ+2,1. The unity input 304 is added to the output of multiplier 313, and to the second intermediate value γ+2,1, in adder 324, to produce a third intermediate value γ+3,1. The first intermediate value γ+1,1 is multiplied by -λ+3 in multiplier 326. The second intermediate value γ+2,1 is multiplied by 2 in multiplier 328 and the product is added to the output of multiplier 326, in adder 330, to produce the third intermediate value γ+3,2. This processing structure is repeated, feeding the unity values into the side of the sparse superlattice structure, until the seventh-order values γ+7,i are produced.
If desired, the linear predictor coefficients can be computed directly from the unit side-fed structure of FIG. 11. FIG. 12 illustrates a simple processing structure for producing the linear predictor a, for a system having order 6, from the sixth intermediate values γ6,i, where i equals 1 through 6 ( γ ⁺6,6 = γ ⁺6,1; - γ ⁺6,5 = γ ⁺6,2,
and the fifth intermediate values γ5,j, where j equals 1 through 5, produced from a side-fed sparse superlattice such as that shown in FIG. 11. Specifically, 6 multipliers 331 through 336 are provided and function to multiply the input value by the quantity -(1 + k₆), which is produced in accordance with equations (36) or (37). The value applied to the first multiplier 331 is unity, and the values applied to the remaining multipliers 332 through 336 are γ+5,1 through γ ⁺5,5, respectively. Six adders 341 through 346 are each associated with one of the multipliers 331 through 336. The first adder 341 receives the output of the first multiplier 331, γ+ 6,1 and unity. These values are added in adder 341 to produce the first linear prediction coefficient a_6,1. The outputs of multiplier 332, adder 341 and γ+ 6,2 are added in adder 342 to produce the linear prediction coefficient a₆,₂. The remaining adders 343 through 346 each receive the output of the previous adder along with the output of the associated multiplier and associated value of γ+ 6,i to thereby yield the linear prediction coefficients a6,3 through a₆,₆.
FIG. 13 illustrates another unit side-fed sparse superlattice, which utilizes the values generated by the subtractive sparse superlattice of FIG. 7, to produce the constant group delay predictor γ―, as follows:
With reference to FIG. 13, the unit side-fed sparse superlattice shown therein includes seven unity inputs 351-357 each of which with the exception of the upper-most input 357, is applied to the input of a respective multiplier 361 through 366. The first multiplier 361 functions to multiply the unity input by the value-λ₂ , produced by the subtractive sparse superlattice of FIG. 7. The output of the multiplier 361 is added to the value γ_1,1 and the unity input at 352, in adder 367, to produce a second intermediate value y2,₁. The output from multiplier 362 is added to the unity input 353, and the output of adder 367, in adder 368, to thereby produce a third intermediate value γ 3,1. The output of adder 367 is also applied to multiplier 369 which functions to multiply the value input thereto by -À4 , and the product is applied to adder 370, which also receives the output of adder 368, and the value γ 3,2 , to thereby produce the sum, y-4,2. This process continues in a similar manner, as shown, until the constant group delay prediction coefficients γ7,1 through γ―7,4 are produced. The values of γ―1,1. γ―3,2, γ―5,3 and γ 7,4 are equal to zero, but are shown in order to illustrate the natural constraint of the constant group delay problem.
In addition to the production of the linear phase predictor and the constant group delay predictor, the sparse superlattice structures, in accordance with the present invention, can be used for optimal LS-FIR filtering, in a manner similar to that described in EP-A-0 282 799, referred to above. With reference to FIG. 14 a superladder structure, called the "sparse superladder", for the production of lattice filter coefficients, in conjunction with the additive sparse superlattice of FIG. 6, is shown. The sparse superladder receives autocorrelation coefficients ri through r₅ at inputs 401 through 405, and crosscorrelation coefficients di through d₆ are applied to input terminals 411 through 416. In addition to the autocorrelation and crosscorrelation coefficients, the sparse superladder receives a number of inputs from the additive sparse superlattice of FIG. 6, specifically, in the case of a system having order 6, as shown in FIG. 14, the first intermediate values of the additive sparse superlattice, amely through
through
, at input terminals 421 through 424, the second intermediate values from the additive sparse superlattice, ²1 , through ²3 at input terminals 431 through 433, third intermediate values ³1 and ³2, at input terminals 441 and 442, and fourth intermediate value ⁴1 , at input terminal 450.
Multipliers 451 through 455 each receive an associated autocorrelation coefficient ri through rs. Similarly, the four first intermediate values ¹1 through ¹4 are applied to respective multipliers 461 through 464, the second intermediate values ²1 through ²―3 are applied to respective multipliers 471, 472 and 473, third intermediate values ³1 and ³―2 are applied to multipliers 474 and 475, respectively, and finally, the fourth intermediate value ⁴1 is applied to multiplier 476. The multipliers in the sparse supperladder function to multiply the input quantities by one of the sparse superlattice filter coefficients c 1 , defined as follows:
The sparse superladder of FIG. 14 is arranged so that first intermediate values ω 0 through ω¹―5, second intermediate values ω0² through ω²-4, third intermediate values ω 0 through ω³-3, fourth intermediate values ω0⁴ through ω⁴-2, fifth intermediate values ω5 0 through ω⁵-1, and sixth intermediate value ω0⁷, are produced in accordance with the following relationship:
The sparse lattice filter coefficients
are applied to the multipliers, as shown. Specifically, filter coefficient -Xc , is applied to multipliers 451 through 455, filter coefficient -λC 2 is applied to multipliers 461 through 464, etc., for the production of the subsequent sparse lattice filter coefficients in accordance with the above equation.
Although explained in general terms, the specific implementations of the structures shown in FIGS. 11-14, in either a fully sequential, fully parallel, or partitioned parallel manner, will be apparent to those skilled in the art in view of the fore-going description.
The additive and subtractive sparse superlattice structures of FIGS. 6 and 7, along with their ancillary structures, shown in FIGS. 11-14, are suitable for efficient hardware implementation. A small number of processors, for example three to six, each containing an adder and a multiplier, may be used for many applications with very good results and a substantial increase in processing speed over that of prior art signal processors.
It will be appreciated that the present invention allows the use of a feasible number of parallel processors to perform highly efficient signal processing. When so implemented, the parallel partitioned implementation provides the benefits of parallel processing, manageable hardware complexity, and optimal signal processing for a given number of available processors. Furthermore, the use of the sparse structures ensures minimum hardware complexity.

Claims

1. A signal processor for receiving autocorrelation coefficients r_;, corresponding to a signal having order p, and for producing quantities λ_i ⁺, defined as (1 - k_i)(k_i-1 + 1), where the quantities k_; correspond to the lattice predictor coefficients, or "PARCORS", of said signal, comprising:

at least one adder (60) for adding the zeroth and first order autocorrelation coefficients to produce a first sum;

a dividers (81) forming the quotient of said first sum and a value related to said zeroth order autocorrelation coefficient, to produce a first of said quantities λ_i ⁺;

at least one multiplier (71) multiplying said first order autocorrelation coefficient by the first of said quantities X_i ⁺, to produce a product;

an adding structure (61, 91) for adding the first and second order autocorrelation coefficients to produce a second sum, and for adding the first sum, the product, and the second sum, to produce a third sum; and

a dividing structure (82) dividing the third sum by the first sum to thereby produce a second of said quantities λ_i ⁺.

2. The signal processor of claim 1 wherein said adding structure comprises a retrieval structure for applying said first sum, said second sum and said product to said at least one adder, to thereby produce said third sum.

3. The signal processor of claim 2 wherein said adding structure comprises an input structure for applying the first and second autocorrelation coefficients to said at least one adder, to thereby produce said second sum.

4. The signal processor of claim 3 further comprising a storage device for storing the first sum.

5. A signal processor for receiving autocorrelation coefficients r_;, corresponding to a signal having order p, and for producing quantities λ_i ⁺, defined as (1-k_i)(k_i-1 + 1), where the quantities k_; correspond to the lattice predictor coefficients, or "PARCORS", of said signal, comprising:

a plurality of adders (201-203);

a plurality of multipliers (204-206)

a storage and retrieval structure (216; 220) for (i) selectively applying a first plurality of pairs of adjacent autocorrelation coefficients to said plurality of adders, to produce an associated plurality of first intermediate values, for (ii) selectively applying a plurality of autocorrelation coefficients to said plurality of multipliers to produce an associated plurality of first products, and for (iii) selectively re-applying an adjacent pair of said first intermediate values, and a selected one of said first products, to each of a plurality of said adders, to produce an associated plurality of second intermediate values; and

a divider (222) for initially dividing a selected first intermediate value by a value related to the zeroth order autocorrelation coefficient, to thereby produce a first of said quantities λ_i ⁺, and thereafter dividing a selected second intermediate value by said selected first intermediate value, to thereby produce a second of said quantities λ_i ⁺;

said first of said quantities λ_i ⁺ being applied to said plurality of multipliers to produce said plurality of second intermediate values.

6. The signal processor of claim 5, wherein said storage and retrieval structure (i) selectively re-applies at least one of said first intermediate variables to at least one of said multipliers, to produce at least one second product, and (ii) selectively re-applies at least one adjacent pair of said second intermediate values, and said second product, to one of said adders, to thereby produce at least one third intermediate value, and wherein said divider divides said third intermediate value by said selected second intermediate value, to thereby produce a third of said quantities λ_i ⁺

7. The signal processor of claim 6, wherein the number of said adders is less than the signal order m, and wherein a second plurality of pairs of adjacent auto-correlation coefficients, less than or equal to, in quantity, said first plurality of pairs of adjacent autocorrelation coefficients, are applied to said plurality of adders subsequent to the generation of said plurality of second intermediate values.

8. The signal processor of claim 5 further including structure for producing linear phase prediction coefficients γ+ i,j of the linear phase predictor y⁺, where γ⁺ = a + Ja, a is the linear predictor, and J is defined as

9. The signal processor of claim 5 further including structure for providing FIR filter coefficients

where

ω0 i, i = -1 to -m, correspond to cross-correlation coefficients di - d_m, and _i ^m correspond to intermediate values of order m.

10. A signal processor for receiving autocorrelation coefficients r_;, corresponding to a signal having order p, and for producing quantities λ_i―, defined as (1 +k_i) (1-k_i-1), where the quantities k_i correspond to the lattice predictor coefficients, or "PARCORS", of said signal comprising:

at least one adder (120) for adding (i) the difference (ro-ri) between the zeroth and first order autocorrelation coefficients, and (ii) the difference (r_i-r₂) between the first and second order autocorrelation coefficients, to produce a first sum;

a divider (141) for forming the quotient of said first sum and the difference between said zeroth and first autocorrelation coefficients, to produce the first of said quantities λ_i―;

at least one multiplier (131) for multiplying the difference (r₁-r₂) between the first and second order autocorrelation coefficients by the first of said quantities λ_i―, to produce a product;

an adding structure (121, 151) for adding (i) the difference (ri-r₂) between the first and second autocorrelation coefficients and (ii) the difference (r₂-r₃) between the second and third autocorrelation coefficients, to produce a second sum, and for adding the first sum, the product, and the second sum, to produce a third sum; and

a dividing structure (142) for dividing said third sum by said first sum, to thereby produce a second of said quantities λ_i―.

11. The signal processor of claim 10 wherein said adding structure comprises a retrieval structure for applying said first sum, said second sum and said product to said at least one adder, to thereby produce said third sum.

12. The signal processor of claim 11 wherein said adding structure comprises an input structure for applying the difference between the first and second autocorrelation coefficients and the difference between the second and third autocorrelation coefficients to said at least one adder, to thereby produce said second sum.

13. The signal processor of claim 12 further comprising a storage device for storing the first sum.

14. A signal processor for receiving autocorrelation coefficients r_;, corresponding to a signal having order p, and for producing quantities λ_i―, defined as (1 + k_;) (1-k,-_i), where the quantities k_; correspond to the lattice predictor coefficients, or "PARCORS", of said signal, comprising

a plurality of adders (201-203)

a plurality of multipliers (204-206)

a storage and retrieval structure (216, 220) for selectively applying a first plurality of pairs of differences between adjacent autocorrelation coefficients to said plurality of adders to produce an associated plurality of second intermediate values, for (ii) selectively applying a plurality of differences between adjacent autocorrelation coefficients to said plurality of multipliers to produce an associated plurality of first products and for (iii) selectively re-applying an adjacent pair of said second intermediate values, and a selected one of said first products, to each of a plurality of said adders, to produce an associated plurality of third intermediate values and

a divider (222) for initially dividing a selected second intermediate value by the difference between the zeroth and first autocorrelation coefficients, to thereby produce a first of said quantities λ_i―, and thereafter dividing a selected third intermediate value by said selected second intermediate value, to thereby produce a second of said quantities λ_i―;

said first of said quantities λ_i― being applied to said plurality of multipliers to produce said plurality of third intermediate values.

15. The signal processor of claim 14, wherein said storage and retrieval structure (i) selectively re-applies at least one of said second intermediate variables to at least one of said multipliers, to produce at least one second product, and (ii) selectively re-applies at least one adjacent pair of third intermediate values, and said second product, to one of said adders, to thereby produce at least one fourth intermediate value, and wherein said divider divides said fourth intermediate value by said selected third intermediate value, to thereby produce a third of said quantities λ_i―.

16. The signal processor of claim 15, wherein the number of said adders is less than the signal order p, and wherein a second plurality of differences between pairs of adjacent autocorrelation coefficients, less than or equal to, in quantity, said first plurality of differences between pairs of adjacent autocorrelation coefficients, are applied to said plurality of adders subsequent to the generation of said plurality of third intermediate values.

17. The signal processor of claim 14 further including structure for producing constant group delay prediction coefficients γi,j of the linear phase predictor γ―, where y = a - Ja, a is the linear predictor and J is defined as